AbstractLoss.h

Go to the documentation of this file.

//===========================================================================
/*!
 * 
 *
 * \brief       super class of all loss functions
 * 
 * 
 *
 * \author      T. Glasmachers
 * \date        2010-2011
 * \file
 *
 * \par Copyright 1995-2017 Shark Development Team
 * 
 * <BR><HR>
 * This file is part of Shark.
 * <https://shark-ml.github.io/Shark/>
 * 
 * Shark is free software: you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as published 
 * by the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 * 
 * Shark is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU Lesser General Public License for more details.
 * 
 * You should have received a copy of the GNU Lesser General Public License
 * along with Shark.  If not, see <http://www.gnu.org/licenses/>.
 *
 */
 
#ifndef SHARK_OBJECTIVEFUNCTIONS_LOSS_ABSTRACTLOSS_H
#define SHARK_OBJECTIVEFUNCTIONS_LOSS_ABSTRACTLOSS_H
 
#include <shark/ObjectiveFunctions/AbstractCost.h>
#include <shark/LinAlg/Base.h>
#include <shark/Core/Traits/ProxyReferenceTraits.h>
namespace shark {
    
/// \defgroup lossfunctions Loss Functions
/// \brief Loss functions define loss values between a model prediction and a given label.
 
/// \brief Loss function interface
///
/// \par
/// In statistics and machine learning, a loss function encodes
/// the severity of getting a label wrong. This is am important
/// special case of a cost function (see AbstractCost), where
/// the cost is computed as the average loss over a set, also
/// known as (empirical) risk.
///
/// \par
/// It is generally agreed that loss values are non-negative,
/// and that the loss of correct prediction is zero. This rule
/// is not formally checked, but instead left to the various
/// sub-classes.
///
/// \ingroup lossfunctions
template<class LabelT, class OutputT = LabelT>
class AbstractLoss : public AbstractCost<LabelT, OutputT>
{
public:
    typedef AbstractCost<LabelT, OutputT> base_type;
    typedef OutputT OutputType;
    typedef LabelT LabelType;
    typedef RealMatrix MatrixType;
 
    typedef typename Batch<OutputType>::type BatchOutputType;
    typedef typename Batch<LabelType>::type BatchLabelType;
 
    /// \brief Const references to LabelType
    typedef typename ConstProxyReference<LabelType const>::type ConstLabelReference;
    /// \brief Const references to OutputType
    typedef typename ConstProxyReference<OutputType const>::type ConstOutputReference;
 
    AbstractLoss(){
        this->m_features |= base_type::IS_LOSS_FUNCTION;
    }
 
    /// \brief evaluate the loss for a batch of targets and a prediction
    ///
    /// \param  target      target values
    /// \param  prediction  predictions, typically made by a model
    virtual double eval( BatchLabelType const& target, BatchOutputType const& prediction) const = 0;
 
    /// \brief evaluate the loss for a target and a prediction
    ///
    /// \param  target      target value
    /// \param  prediction  prediction, typically made by a model
    virtual double eval( ConstLabelReference target, ConstOutputReference prediction)const{
        BatchLabelType labelBatch = Batch<LabelType>::createBatch(target,1);
        getBatchElement(labelBatch,0)=target;
        BatchOutputType predictionBatch = Batch<OutputType>::createBatch(prediction,1);
        getBatchElement(predictionBatch,0)=prediction;
        return eval(labelBatch,predictionBatch);
    }
 
    /// \brief evaluate the loss and its derivative for a target and a prediction
    ///
    /// \param  target      target value
    /// \param  prediction  prediction, typically made by a model
    /// \param  gradient    the gradient of the loss function with respect to the prediction
    virtual double evalDerivative(ConstLabelReference target, ConstOutputReference prediction, OutputType& gradient) const {
        BatchLabelType labelBatch = Batch<LabelType>::createBatch(target,1);
        getBatchElement(labelBatch, 0) = target;
        BatchOutputType predictionBatch = Batch<OutputType>::createBatch(prediction, 1);
        getBatchElement(predictionBatch, 0) = prediction;
        BatchOutputType gradientBatch = Batch<OutputType>::createBatch(gradient, 1);
        double ret = evalDerivative(labelBatch, predictionBatch, gradientBatch);
        gradient = getBatchElement(gradientBatch, 0);
        return ret;
    }
    
    /// \brief evaluate the loss and its first and second derivative for a target and a prediction
    ///
    /// \param  target      target value
    /// \param  prediction  prediction, typically made by a model
    /// \param  gradient    the gradient of the loss function with respect to the prediction
    /// \param  hessian     the hessian of the loss function with respect to the prediction
    virtual double evalDerivative(
        ConstLabelReference target, ConstOutputReference prediction,
        OutputType& gradient,MatrixType & hessian
    ) const {
        SHARK_FEATURE_EXCEPTION_DERIVED(HAS_SECOND_DERIVATIVE);
        return 0.0;  // dead code, prevent warning
    }
 
    /// \brief evaluate the loss and the derivative w.r.t. the prediction
    ///
    /// \par
    /// The default implementations throws an exception.
    /// If you overwrite this method, don't forget to set
    /// the flag HAS_FIRST_DERIVATIVE.
    /// \param  target      target value
    /// \param  prediction  prediction, typically made by a model
    /// \param  gradient    the gradient of the loss function with respect to the prediction
    virtual double evalDerivative(BatchLabelType const& target, BatchOutputType const& prediction, BatchOutputType& gradient) const
    {
        SHARK_FEATURE_EXCEPTION_DERIVED(HAS_FIRST_DERIVATIVE);
        return 0.0;  // dead code, prevent warning
    }
    
    /// from AbstractCost
    ///
    /// \param  targets      target values
    /// \param  predictions  predictions, typically made by a model
    double eval(Data<LabelType> const& targets, Data<OutputType> const& predictions) const{
        SIZE_CHECK(predictions.numberOfElements() == targets.numberOfElements());
        SIZE_CHECK(predictions.numberOfBatches() == targets.numberOfBatches());
        int numBatches = (int) targets.numberOfBatches();
        double error = 0;
        SHARK_PARALLEL_FOR(int i = 0; i < numBatches; ++i){
            double batchError= eval(targets.batch(i),predictions.batch(i));
            SHARK_CRITICAL_REGION{
                error+=batchError;
            }
        }
        return error / targets.numberOfElements();
    }
 
    /// \brief evaluate the loss for a target and a prediction
    ///
    /// \par
    /// convenience operator
    ///
    /// \param  target      target value
    /// \param  prediction  prediction, typically made by a model
    double operator () (LabelType const& target, OutputType const& prediction) const
    { return eval(target, prediction); }
 
    double operator () (BatchLabelType const& target, BatchOutputType const& prediction) const
    { return eval(target, prediction); }
 
    using base_type::operator();
};
 
 
}
#endif